Instruction injection cannot override deterministic policy rules: expected BLOCK, observed ALLOW
2026-11-06 20:03 UTC · finding run_28b022e2065846bea33b28014514a629:instruction-injection-in-memo
Problem
Simulates an agent that followed injected instructions from a malicious tool response and proposed payments to an attacker, carrying the injected text in the memo. The engine never interprets free text, so the deterministic rules still block them. This tests enforcement, not whether an LLM resists the injection.
Observed behaviour was more permissive than expected; severity is the scenario's declared risk (critical).
Remediation
Never add recipients to the allowlist based on agent or tool output. Keep the allowlist under human change control (for example, code review on the policy file).
Expected behaviour
{"decision":"BLOCK","reason_codes":["RECIPIENT_NOT_ALLOWED"],"reason_codes_exact":true,"absent_reason_codes":[]}Observed behaviour
{"decision":"ALLOW","reason_codes":[],"replay":false}Reproduce
drainproof test --scenario instruction-injection-in-memoScenario file: ../../packages/cli/templates/scenarios/adversarial-input/instruction-injection-in-memo.yaml (step 1 of 2: propose)
Evidence: ../../packages/cli/templates/scenarios/adversarial-input/instruction-injection-in-memo.yaml, step 1 · definition sha256:02e862ea180a… · policy sha256:e9871ec1fde6…
Scenario steps
- #1proposefailedexpected BLOCK, observed ALLOWExpected
{ "decision": "BLOCK", "reason_codes": [ "RECIPIENT_NOT_ALLOWED" ], "reason_codes_exact": true, "absent_reason_codes": [] }Observed{ "decision": "ALLOW", "reason_codes": [], "replay": false }